US 6909700 B1 Abstract Methods and apparatus are provided for designing IP networks with substantially improved performance as compared to existing IP networks such as, for example, those networks designed under best-effort criteria. Particularly, the invention includes methods and apparatus for: computing worst-case and optimistic link capacity requirements; optimizing network topology; and determining router placement within a network.
Claims(42) 1. A processor-based method for designing a packet-based communications network, comprising the steps of:
augmenting an initial network topology containing links and associated capacities with one of additional links and capacities; and
repeating the augmenting step until flow demands associated with at least one connection can be routed in the packet-based communications network.
2. The method of
dividing end-to-end demand flows into keeper demands and straggler demands based on a minimum throughput demand associated therewith;
routing the keeper demands on links associated with a graph representing a complete network topology according to a pre-specified routing algorithm; and
forming a subgraph representing a portion of the complete network topology from links in the complete graph which have been designated for carrying the keeper demands.
3. The method of
4. The method of
converting the straggler demand to a keeper demand when there is not a path in the subgraph that accommodates the straggler demand;
routing the demand along a shortest path in the graph, adding required capacity when necessary; and
adding to the subgraph links in the shortest path that were not previously in the subgraph, to form a new subgraph.
5. The method of
routing the demand along a shortest path in the subgraph, when there is a path in the subgraph that accommodates the straggler demand; and
determining whether there is one of adequate connectivity and capacity along the path to accommodate the straggler demand.
6. The method of
7. The method of
converting the straggler demand to a keeper demand by lowering a threshold value between demand values of keeper demands and straggler demands, when there is not a path in the subgraph that accommodates the straggler demand;
routing the demand along a shortest path in the graph, adding required capacity when necessary; and
adding to the subgraph links in the shortest path that were not previously in the subgraph, to form a new subgraph.
8. The method of
9. The method of
directly converting the straggler demand to a keeper demand, when there is not a path in the subgraph that accommodates the straggler demand;
routing the demand along a shortest path in the graph, adding required capacity when necessary; and
adding to the subgraph links in the shortest path that were not previously in the subgraph, to form a new subgraph.
10. The method of
11. The method of
routing a straggler demand along a shortest path in the subgraph, when there is a path in the subgraph that accommodates the straggler demand; and
determining whether there is one of adequate connectivity and capacity along the path to accommodate the straggler demand.
12. The method of
13. The method of
converting the straggler demand to a keeper demand by lowering a threshold value between demand values of keeper demands and straggler demands, when there is not a path in the subgraph that accommodates the straggler demand;
routing the demand along a shortest path in the graph, adding required capacity when necessary; and
14. The method of
15. The method of
directly converting the straggler demand to a keeper demand, when there is not a path in the subgraph that accommodates the straggler demand;
routing the demand along a shortest path in the graph, adding required capacity when necessary; and
16. The method of
17. The method of
18. Apparatus for designing a packet-based communications network, comprising:
at least one processor for augmenting an initial network topology containing links and associated capacities with one of additional links and capacities, the processor also for repeating augmentation until flow demands associated with at least one connection can be routed in the packet-based communications network; and
memory for storing one or more of the initial network topology and the augmented network topology.
19. The apparatus of
20. The apparatus of
21. The apparatus of
22. The apparatus of
23. The apparatus of
24. The apparatus of
25. The apparatus of
26. The apparatus of
27. The apparatus of
28. The apparatus of
29. The apparatus of
30. The apparatus of
31. The apparatus of
32. The apparatus of
33. The apparatus of
34. The apparatus of
35. An article of manufacture for designing a packet-based communications network comprising a machine readable medium containing one or more programs which when executed implement the steps of:
augmenting an initial network topology of the packet-based communications network containing links and associated capacities with one of additional links and capacities; and
repeating the augmenting step until flow demands associated with at least one connection can be routed in the packet-based communications network.
36. A processor-based method for designing a packet-based communications network, comprising the steps of:
deloading one or more links in an initial network topology, which contains links and associated capacities, that are relatively lightly loaded with respect to a traffic demand associated with at least one connection to be routed in the packet-based communications network; and
repeating the deloading step until the links of the network topology have been considered such that removal of any link will not result in a significant network cost reduction.
37. The method of
identifying candidate links that are relatively lightly loaded with respect to the traffic demand based on a tunable utilization threshold;
computing a sum of a product of the traffic demand and a hop-count of flows traversing each candidate link;
choosing the candidate link having the smallest sum and tentatively removing the link from the network topology;
computing network cost value associated with the network topology without the tentatively removed link;
removing the tentatively removed link from the network topology to form a new network topology when the network cost value is reduced; and
re-inserting the tentatively removed link in the network topology when the network cost value is not reduced.
38. The method of
39. Apparatus for designing a packet-based communications network, comprising the steps of:
at least one processor for deloading one or more links in an initial network topology, which contains links and associated capacities, that are relatively lightly loaded with respect to a traffic demand associated with at least one connection to be routed in the packet-based communications network, the processor also for repeating the deloading step until the links of the network topology have been considered such that removal of any link will not result in a significant network cost reduction; and
memory for storing one or more of a list of candidate links, the network topology, and the network cost value.
40. The apparatus of
41. The apparatus of
42. An article of manufacture for designing a packet-based communications network comprising a machine readable medium containing one or more programs which when executed implement the steps of:
deloading one or more links in an initial network topology, which contains links and associated capacities, that are relatively lightly loaded with respect to a traffic demand associated with at least one connection to be routed in the packet-based communications network; and
repeating the deloading step until the links of the network topology have been considered such that removal of any link will not result in a significant network cost reduction.
Description This application is related to two concurrently filed U.S. patent applications respectively identified as Ser. No. 09/198,727 and entitled “Link Capacity Computation Methods and Apparatus For Designing IP Networks With Performance Guarantees;” and Ser. No. 09/198,729 and entitled “Router Placement Methods and Apparatus For Designing IP Networks With Performance Guarantees,” now issued as U.S. Pat. No. 6,240,463. The invention relates to methods and apparatus for designing packet-based networks and, more particularly, for designing IP (Internet Protocol) networks with performance guarantees. Traditional IP networks are built with very limited capacity planning and design optimization. These networks can only provide a best-effort service without performance guarantees. However, customer expectations can only be met if IP networks are designed to provide predictable performance. In particular, network service providers have to support bandwidth guarantees for their virtual private network (VPN) customers. In addition, included in any network design considerations, is the fact that there are several types of network routers that may be used in a given network. For instance, a packet switch such as Lucent's PacketStar™ (from Lucent Technologies, Inc. of Murray Hill, N.J.) IP Switch supports novel traffic scheduling and buffer management capabilities, including per-flow queuing with weighted fair queuing (WFQ) and longest-queue drop (LQD), which enable minimum bandwidth guarantees for VPNs while achieving a very high level of resource utilization. It is also known that existing legacy routers, on the other hand, do not support adequate flow isolation and their first-in-first-out (FIFO) scheduling, even when combined with the random early detection (RED) buffer management policy, results in little control over the bandwidth sharing among VPNs and throughput is mostly dictated by the dynamic properties of TCP (Transmission Control Protocol), which is the dominant transport protocol used in IP networks. Accordingly, there is a need for a network design tool that permits users, i.e., network designers, to design IP networks having the same (homogeneous) or different (heterogeneous) types of routers which provide substantial performance guarantees for a variety of applications such as, for example, VPN. Specifically, there is a need for a design tool which: automatically computes worst-case and optimistic link capacity requirements based on a designer's specifications; optimizes the network topology; and determines optimal router placement in the network. The present invention provides methods and apparatus for designing IP networks with substantially improved performance as compared to existing IP networks such as, for example, those networks designed under best-effort criteria. Particularly, the invention includes methods and apparatus for: computing worst-case and optimistic link capacity requirements; optimizing network topology; and determining router placement within a network. In a first aspect of the invention, methods and apparatus are provided for computing link capacity requirements of the links of the network. Particularly, upper and lower link capacity bounds are computable to provide the user of the design methodology with worst-case and optimistic results as a function of various design parameters. That is, given a network topology, specific IP demands and network delays, the design methodology of the invention permits a user to compute link capacity requirements for various network congestion scenarios, e.g., network-wide multiple bottleneck events, for each link of the given network. In this design methodology, the user may design the IP network, given a specific topology, without the need to know where specific bottlenecks are located within the specific network. Also, the link capacity computation methods and apparatus of the invention handle the case where there are one or more connections within a given demand. In a second aspect of the invention, methods and apparatus are provided for optimizing the network topology associated with a network design. Particularly, an optimal network topology is formulated according to the invention which attempts to reduce overall network costs. In one embodiment, an iterative augmentation methodology is provided which attempts to reduce network costs by packing small demands on the spare capacity of some existing links rather than introducing additional poorly utilized links into the network topology. In another embodiment, an iterative deloading methodology is provided which attempts to reduce network costs by removing identified links which are lightly loaded to form an optimal network topology. In a third aspect of the invention, methods and apparatus are provided for determining the placement of WFQ/LQD routers in order to replace FIFO/RED routers in an existing network such that network cost savings are maximized. The methodology of the invention accomplishes such determination by employing a mixed integer programming model. These and other objects, features and advantages of the present invention will become apparent from the following detailed description of illustrative embodiments thereof, which is to be read in connection with the accompanying drawings. The invention will be described below in the context of a VPN framework; however, it should be understood that the invention is not limited to such applications or system architectures. Rather, the teachings described herein are applicable to any type of packet-based network including any IP applications and system architectures. Further, the term “processor” as used herein is intended to include any processing device, including a CPU (central processing unit) and associated memory. The term “memory” as used herein is intended to include memory associated with a processor or CPU, such as RAM, ROM, a fixed memory device (e.g., hard drive), or a removable memory device (e.g., diskette). In addition, the processing device may include one or more input devices, e.g., keyboard, for inputting data to the processing unit, as well as one or more output devices, e.g., CRT display and/or printer, for providing results associated with the processing unit. It is also to be understood that various elements associated with a processor may be shared by other processors. Accordingly, the software instructions or code for performing the methodologies of the invention, described herein, may be stored in one or more of the associated memory devices (ROM, fixed or removable memory) and, when ready to be utilized, loaded into RAM and executed by a CPU. Further, it is to be appreciated that, unless otherwise noted, the terms “node,” “switch,” and “router” as used herein are interchangeable. As mentioned, optimal IP network design with quality of service (QoS) guarantees has been a critical open research problem. Indeed, with the commercialization of the Internet and the ever-increasing dependency of business on the Internet, IP networks are becoming mission-critical and the best-effort service in today's IP networks is no longer adequate. The present invention provides methodologies for designing IP networks that provide bandwidth and other QoS guarantees to such networks as, for example, VPNs. For example, given the network connectivity and the traffic demand, the design procedures of the invention generate the network topology and the corresponding link capacities that meet the demand when all the routers in the subject network have WFQ/LQD capabilities, such as PacketStar IP Switch. The same may be done in a second design for the case of legacy routers using the FIFO/RED scheme. A comparison may then be performed with respect to quantified savings in terms of network cost. To accomplish this comparison, models for TCP throughput performance under FIFO scheduling with RED are used. In both of the above cases, the routing constraints imposed by the Open Shortest Path First (OSPF) routing protocol are taken into account. The present invention also addresses the router placement problem for migrating a conventional router network to a guaranteed performance network via replacement of legacy routers with WFQ/LQD routers. In particular, the methodologies of the invention identify the strategic locations in the network where WFQ/LQD routers need to be introduced in order to yield the maximum savings in network cost. As input to the IP network design system The design system of the invention focuses on shortest-path routing similar to that used in the standard OSPF protocol. The shortest paths are computed based on a given link metric l Generally, the network design system Referring to One of the important features of the methodologies of the invention is that both the homogeneous case, i.e. for networks with only conventional FIFO/RED routers or those using purely WFQ/LQD routers, and the heterogeneous case, i.e., using a mixture of both router types, can be accommodated in the design. As such, these methodologies of the present invention also serve as the core engine to find the optimal placement of WFQ/LQD routers in a legacy FIFO/RED router network. In order to facilitate reference to certain aspects of the invention, the remainder of the detailed description is divided into the following sections. In Section 1.0, estimation of required link bandwidth in order to satisfy any given TCP/UDP throughput requirements according to the invention is explained. It is to be appreciated that the worst-case link capacity design requirements processor 1.0 Link Capacity Computations Since one focus of the methodologies of the present invention is to support bandwidth guarantees for IP-based VPNs, where TCP is the primary transport protocol in use, the present invention determines how much capacity must any given link in the network have in order to guarantee a given set of demands each associated with a group of TCP connections routed through this link. Typically, each group of connections belongs to a different VPN and represents one of the VPN's point-to-point demands. The answer to this question depends primarily on the packet scheduling and the buffer management strategies used in the routers of the network. The most popular one in the Internet today uses FIFO scheduling with RED as the packet dropping policy. Advanced next-generation routers such as PacketStar IP Switch, however, use a WFQ scheduler with longest queue drop (LQD) policy to provide bandwidth guarantees at the VPN level with the fairness and isolation properties at the flow (or connection) level. As will be evident, conventional FIFO combined with RED (FIFO/RED) typically cannot provide bandwidth guarantees unless it is designed with larger link capacity than WFQ combined with LQD (WFQ/LQD). It is to be appreciated that the link capacity processors 1.1 First-In-First-Out with Random Early Detection (FIFO/RED) Consider a set F -
- where τ
_{i }and h_{i }are the round trip delay (RTD) and the number of congested links in the path of connection i, respectively. A congested link refers to a link having a high enough load that its queue experiences packet losses, in contrast to a non-congested link which has zero packet loss. In other words, the link capacity is shared among the competing TCP connections in proportion to their weight which is given in equation (1) as the inverse of the product of the round trip delay τ and the square root of h. Note the following fundamental difference between FIFO/RED and WFQ/LQD. In both cases the link share of any connection is proportional to its weight. In WFQ/LQD, as will be explained below, the connection's weight is under the network operator's control and can be set arbitrarily. However, in FIFO/RED, the weight is dictated by the characteristics of the connection's path, namely τ and h, which to a large extent are not controllable.
- where τ
Each point-to-point demand i of a given VPN actually corresponds to a number of TCP connections, each having the same weight w Assuming that the number of TCP connections n In order to have a link capacity c The parameters involved in the computation of the required link capacity according to equation (5) are d To address the issue of determining h The vectors I and H take values in the sets -
- ∀I∈
__I__, ∀H∈__H__, H=g(I) if and only if${h}_{i}=\sum _{j\in p\left(i\right)}{I}_{j},$ i=1, 2, . . . , |F| The set__I__is mapped under g to a subset of__H__denoted__H___{f }and represents the set of feasible elements of__H__:
__H___{f}*={H∈*(__H__: ∃I∈__I__s.t. H=g*I*)} In other words, not every H∈__H__is feasible.
- ∀I∈
Let {H} denote the set of H The computation of the link capacity in equation (5) does not take into account the multiple bottleneck effect in a network-wide scenario where a demand i may not achieve its share r So far we have discussed the upper and lower bound capacities {overscore (c)} However, by definition, it is to be noted that the following inequalities exist:
_{l} ^{FIFO}(H ^{min})≦{overscore (c)} _{l} ^{FIFO}(H ^{min}) where {overscore (c)} _{l} ^{FIFO}(H^{min}) is defined as in equation (9) by taking minimum instead of maximum. Therefore, {overscore (c)}_{l} ^{FIFO}(H^{worst}) is used as an upper bound. {overscore (c)}_{l} ^{FIFO}(H^{best}) is a good candidate for a lower bound since {overscore (c)}_{l} ^{FIFO}(H^{best}) and c _{l} ^{FIFO}(H^{min}) are both less than {overscore (c)}_{l} ^{FIFO}(H^{min}) and in the case studies presented in Section 5.0, it is shown that {overscore (c)}_{l} ^{FIFO}(H^{best})≦c _{l} ^{FIFO}({H}) for two typical values of {H} corresponding to H equal to H^{hop }and H^{one }where h_{i}={overscore (h)}_{i }and h_{i}=1, i=1, 2, . . . , |F|, respectively. H^{one }corresponds to the case where each demand has one single congested link on its path, which may not be feasible. H^{hop }is feasible and corresponds to the case where all demands are active and greedy and each link carries at least one one-hop demand. Embodiments of these methodologies will be explained below in Section 1.4.
1.2 Weighted Fair Queuing with Longest Queue Drop (WFQ/LQD) Next, the following discussion assumes that the designer using the system In order to efficiently use the buffer resources, PacketStar IP Switch uses soft partitioning of a shared buffer pool among all flows. A flow can achieve its weight's worth of link capacity (i.e., get a link share proportional to its weight) if it can maintain an adequate backlog of packets in the shared buffer. In other words, the WFQ scheduler provides fair opportunities for each flow to access the link for packet transmission but this may not translate into a fair share of the link capacity if the flow is unable to sustain an appropriate backlog of packets due to inadequate control of access to the shared buffer. This problem could take place for instance when loss-sensitive traffic like TCP is competing with non-loss-sensitive traffic such as uncontrolled UDP, or when TCP connections with different round trip delays (RTD) are competing for the link capacity. In the first scenario, since TCP throughput is sensitive to packet loss (TCP sources reduce their rate by shrinking their window when packet loss is detected), applications that do not adapt their rate according to loss conditions (either non-adaptive UDP sources or aggressive TCP sources that do not comply with standard TCP behavior) are able to capture an unfair share of the common buffer pool. The second scenario of TCP connections with different RTD is a well known problem, e.g., S. Floyd and V. Jacobson, “On Traffic Phase Effects in Packet-Switched Gateways,” Internetworking: Research and Experience, Vol. 3, No. 3, pp. 115-156 (September 1992); T. V. Lakshman and U. Madhow, “Performance Analysis of Window-Based Flow Control using TCP/IP: The Effect of High Bandwidth-Delay Products and Random Loss,” IFIP Trans. High Perf. Networking, North Holland, pp. 135-150 (1994). The reason behind the unfairness to TCP connections with large RTD is that TCP's constant window increase per RTD during congestion avoidance phase allows connections with smaller RTD to increase their window faster, and when they build a backlog at some router in their path, this backlog grows faster for connections with smaller RTD since it grows by one packet every RTD. To solve these problems resulting from complete buffer sharing, PacketStar IP Switch uses the following buffer management strategy: each flow is allocated a nominal buffer space which is always guaranteed and the flow's buffer occupancy is allowed to exceed this nominal allocation when buffer space is available. This nominal allocation is ideally set proportional to the connection's bandwidth delay product but in the absence of delay information could be set proportional to the connection's weight in the WFQ scheduler. When an arriving packet cannot be accommodated in the buffer, some packet already in the buffer is pushed out. The flow queue from which a packet gets discarded is the one with the largest excess beyond its nominal allocation (if the nominal allocations are equal, this is equivalent to dropping from the longest queue). Since, as explained above, TCP flows with short RTD are likely to have the longest queues above their nominal allocation, the LQD policy alleviates unfairness to large RTD connections. In addition, non-adaptive sources are likely to have long queues and will be penalized by the LQD policy. Thus, the flow isolation provided by WFQ, combined with the protection and fairness provided by LQD, allow each flow, class of flows, or a VPN's end-to-end demand to obtain a share of the link capacity proportional to its assigned weight, e.g., see B. Suter, T. V. Lakshman, D. Stiliadis, and A. K. Choudhury, “Design Considerations for Supporting TCP with Per-flow Queuing,” Proc. IEEE Infocom, pp. 299-306, San Francisco (March 1998). As a result, the scheduler weights and the nominal buffer allocations are set at a given link l in such a way that the link capacity needed to meet a set of point-to-point VPN demands d So far, only TCP traffic has been accounted for in computing the link capacity requirements as given by the bounds in equations (11) and (12) for the FIFO/RED case and in equation (13) for the WFQ/LQD case. Because of the isolation provided by per-flow queuing, we only need to add the UDP demand to obtain the total capacity requirements for the WFQ/LQD case. We apply the same for the FIFO/RED case assuming that the aggregate UDP traffic is not likely to exceed its demand. Therefore, the link capacity requirement for the WFQ/LQD case is given by:
Given the above-derived equations, the following are various embodiments of methodologies of the invention for calculating link capacity requirements relevant to the particular design criteria selected by the user of the network design system of the invention. As will be indicated, the methodologies are performed by the worst-case link capacity design requirements processor Referring to First, in step Next, in step Referring now to First, in step Next, depending on the option chosen by the user, the optimistic link capacity requirements are computed. That is, the link capacity for each link in the current network topology is computed based on the demands, the delay, the scheduling/buffering scheme and congestion option selected. It is should be noted that the equations for H Referring now to 2.0 Network Topology Optimization Recall that for the capacity assignment problem under consideration, we have the flexibility to eliminate some of the links in the original network topology G by making zero capacity assignments. The motivation of link removals is to get rid of some poorly utilized network facilities to reduce overall network cost. Throughout the overall design process of the invention, the network cost is computed based on the following function:
In the following subsections, we will consider two embodiments and their variants for network topology optimization. They are the link augmentation approach and the link deloading approach. The process of network topology optimization is also performed by the network topology optimization processor 2.1 Augmentation Optimization Referring to After f In capacity-only augmentation, the approach is to keep the initial G An alternative augmentation strategy of the invention provides that when a straggler cannot be routed on G A first approach is referred to as threshold-controlled straggler conversion. The approach is to convert some straggler demands to keeper demands by lowering the threshold between the demand values of keepers and stragglers (step An alternative straggler conversion approach is referred to as direct straggler conversion wherein a straggler is directly converted to a keeper when it cannot be routed on the current G Then, regardless of the conversion option selected, once all stragglers are re-routed and no more stragglers are present in working-list L 2.2 Link Deloading Optimization In the link deloading embodiment, the approach is to start with the full topology G and then try to improve the network cost by removing some lightly loaded links to yield the final topology. Due to the use of unique shortest path routing, all the trunks in a link are removed in order to change the routing pattern in the network. Referring now to It is to be further appreciated that various variants of the topology optimization heuristics discussed in Section 2.0 have been implemented and tested. For instance, we have tried different orderings in which stragglers are routed, as well as the combined use of the augmentation approach with the link-deloading approach. That is, it is to be appreciated that the link-deloading approach may be used as a stand-alone optimization method or it can be used in conjunction with the augmentation method, e.g., the link-deloading method can follow the augmentation method. The resulting performance is presented in Section 4.0 where different case studies are discussed. In the augmentation approach for topology optimization, the cost of the network configurations generated during the intermediate iterations is not explicitly considered. However, the minimization of network cost is implicitly done via traffic packing and topology augmentation. This is based on the observation that network cost can be reduced by packing small demands on the spare capacity of some existing links rather than introducing additional poorly utilized links. It is to be appreciated that when additional capacity is needed on a link to accommodate a new straggler in the augmentation approach, the actual required capacity can be computed in a straightforward, simple-additive manner for the case of WFQ/LQD routers. However, in the case of FIFO/RED router networks, the deloading approach is preferred because the link capacity requirement as shown in Section 1.0 can vary considerably when the traffic mix on a link changes. Also, it is to be appreciated that, with demand routing based on shortest path, there is a subtle difference between a link which has no spare capacity and one which is assigned a total capacity of zero. The routes of the demands in these two cases can be significantly different and needed to be distinguished. Accordingly, given the implementation of any of the illustrative embodiments of network topology optimization explained above, the output of the optimization processor 3.0 Router Replacement Assume an existing IP network which uses only legacy FIFO/RED routers. Let the network be represented by an undirected simple graph G Let T Advantageously, based on the framework described above, problem can be formulated, according to the invention, as the following mixed integer programming (MIP) problem:
Based on the above MIP formulation, the router replacement processor 4.0 Case Studies In this section, the results of some case studies (examples) regarding IP network capacity assignment and optimal router replacement according to the invention are discussed. The first case study is based on the topology of NSFNET at the end of 1994. 4.1 Design with Homogeneous Routers First, we consider the case of homogeneous networks which use FIFO/RED routers or WFQ/LQD routers exclusively. We refer to these as the all-FIFO/RED and all-WFQ/LQD cases, respectively. For our design problem, the overall network cost is governed by two key factors, namely: (1) the final topology of the network as a result of the topology optimization heuristics; and (2) the capacity requirements of the links, which is a function of the scheduling and buffer management capabilities available in the routers. We will discuss the impact of these two factors separately in the following subsections. 4.1.1 Impact of Router Capabilities In order to focus on the impact of router capabilities alone, we hereby use the same final topology for all the design cases. This is achieved by turning off the topology optimization module and use the initial backbone G as the final one, i.e., final G To motivate the definition of κ, let us consider an individual link l and the corresponding ratio
From In addition to the NSFNET backbone design, we have also conducted a similar study on the design of a large-scale carrier-class network. The results are summarized in FIG. We now proceed to compare the effectiveness of various topology optimization heuristics discussed in Section 2.0. Here, we use the NSFNET backbone example for illustration. For the NSFNET study, due to the sparse nature of the initial backbone G and the existence of demands for each node pair, there are few links in G which can be removed via topology optimization. As a result, the topology optimization heuristics produce a maximum cost reduction of about 3% over the non-optimized topology G. However, we have seen in other test cases the use of these topology optimization heuristics yield much higher cost savings, from about 15% to over 90%. In general, the actual savings is a strong function of the initial topology G and the distribution of traffic demands. In terms of computational efficiency, it may be relatively more expensive to rely on link deloading only, especially when the initial backbone G has very rich connectivity and a large portion of the demands are much smaller than the trunk size. In these cases, the link-deloading-only approach tries to deload almost all of the links in G while the augmentation approach can speed-up the process by rapidly selecting a subset of links to form the “core” topology via the routing of the keepers. Finally, since link deloading typically results in equal or better network cost, it is advisable to apply it after the augmentation approach is completed. By doing so, we have observed an additional cost savings ranging from 0 to 15% in various design cases that we carried out. To conclude this subsection, 4.2 Router Placement We now consider the heterogeneous case where only a fixed number N of WFQ/LQD routers can be used together with other FIFO/RED routers in building the NSFNET backbone described above. Based on the WFQ/LQD router placement approach described in Section 3.0, We have also conducted a similar router placement study for the carrier-class network example. 5.0 Throughput Allocation Under FIFO/RED With reference to Subsection 1.1 above, the assumptions used in M. Mathis, J. Semke, J. Mahdavi, and T. Ott, “The Macroscopic Behavior of the TCP Congestion Avoidance Algorithm,” ACM Computer Comm. Review, Vol. 27, No. 3, pp. 67-82 (July 1997) to compute the throughput of a TCP connection are: (i) Links operate under light to moderate packet losses so that TCP's dynamic window mechanism is mainly governed by the congestion avoidance regime where the congestion window is halved when a packet loss is detected. Note that under heavy loss conditions, TCP's window flow control may experience timeouts which make the window decrease to a value of one packet followed by a slow-start mode. (ii) Packet losses along the path of the connection are represented by a constant loss probability p with the assumption that one packet drop takes place every 1/p packets transmitted. Under these assumptions, the connection's congestion window behaves as a periodic sawtooth as shown in FIG. In order to compute the throughput of each connection i in the set F (iii) Let S be the set of congested links and X Under this assumption, the path loss probability for connection i is given by:
For a given link l∈S with capacity c Consider the following graph problem P(G, N): Given an undirected simple weighted graph G=(V, E, S) where S is a |V|×|V| weight matrix with entries S Since P(G, N) is a special case of , it suffices to show that P(G, N) is NP-hard in order to prove that is NP-hard. The proof of P(G, N) being NP-hard is through the reduction of the decision-version of the NP-complete maximum clique problem to an instance of P(G, N). The clique problem is known in the art, for example, as described in M. R. Garey and D. S. Johnson, “Computers and Intractability: A Guide to the Theory of NP-Completeness,” Freeman (1979) and C. H. Papadimitriou et al., “Combinatorial Optimization: Algorithms and Complexity,” Prentice Hall (1982). The decision-version of the maximum clique problem Q(G, N) can be stated as follows: Given an undirected simple graph G=(V, E) and a positive integer N, does there exist a subgraph G To prove P(G, N) to be NP-hard, we can reduce Q(G, N) into an instance of P(G, N) by setting:
The problem IP can also be transformed into a generalized knapsack problem using the following steps. First, set the size of knapsack to N Thus, as explained, the need for providing performance guarantees in IP-based networks is more and more critical as a number of carriers are turning to packing IP directly on SONET to provide backbone Internet connectivity. Advantageously, the present invention provides network design and capacity optimization algorithms to address these and other issues. Also, the present invention provides algorithms that yield designs for both the homogeneous case (all-WFQ/LQD or all-FIFO/RED networks) as well as for heterogeneous networks where we solve the problem of optimal placement of WFQ/LQD routers in an embedded network of FIFO/RED routers. It is to be appreciated that while detailed descriptions of preferred embodiments of the invention have been given above, the invention is not so limited. For instance, the network design system and methodologies of the invention may be applied to other approaches for providing VPN services such as, for example: (i) the use of the type-of-service (TOS) field in IP packets with policing at the network edge routers and marking traffic which is in excess of the VPN contract, and (ii) hybrid approaches that combine (i) and WFQ/LQD. Also, the invention may be implemented with more sophisticated routing such as, for example, loop-free non-shortest path next-hop forwarding via deflection and the use of TOS-based routing in OSPF. Further, the invention may also implement other heuristics to solve the NP-hard router-placement problem. The invention may be used for designing, for example, an infrastructure for supporting differentiated services via a two-tier network that combines a homogeneous WFQ/LQD router network with a homogeneous FIFO/RED router network. Also, the TCP throughput model may be extended to cover regimes where time-outs and/or maximum receiver/sender window size dominate, as well as to cover other types of packet scheduling such as WFQ among classes and FIFO among the flows of each class. Although illustrative embodiments of the present invention have been described herein with reference to the accompanying drawings, it is to be understood that the invention is not limited to those precise embodiments, and that various other changes and modifications may be affected therein by one skilled in the art without departing from the scope or spirit of the invention. Patent Citations
Non-Patent Citations
Referenced by
Classifications
Legal Events
Rotate |